Making GenAI revision accountable: learner evaluation, uptake, and revision quality in source-based academic writing

Generative artificial intelligence (GenAI) can provide fluent revision suggestions, but source-based writing requires learners to judge whether those suggestions preserve source meaning, evidence, attribution, task requirements, and intended stance. This classroom-based study examined 40 second-year English majors completing two source-based writing tasks. The dataset comprised 160 drafts, 80 reflections, 793 traceable revision episodes, and independent ratings of 80 D1–D2 pairs. Cross-classifying 131 reflection-linked rationales with uptake showed that language-clarity reasons were concentrated in local language uptake, trust or uncertainty frequently accompanied checked uptake, and efficiency or time was concentrated in weakly checked uptake. Excluding 34 risk-unclear episodes, substantive-risk suggestions were more often associated with visible checking than local-only suggestions (32.5% vs. 18.9%; participant-clustered GEE, OR = 2.07, 95% CI [1.48, 2.91], p < 0.001), although the authorial-stance contrast was not robust. Within substantive-risk episodes with determinate consequences, visible checking was associated with higher odds of local improvement (84.6% vs. 39.6%; adjusted OR = 8.11, 95% CI [4.88, 13.47], p < 0.001) and lower odds of degradation. D2 texts were rated higher than D1 texts, but absolute rater agreement was moderate, ICC(A,k) = 0.625, and no participant–task-level uptake proportion was significantly associated with whole-text gain after multiplicity correction. The findings are consistent with risk-calibrated accountability as an observational design principle; they do not establish a causal effect of visible checking.